Attention Module Distribution in the MobileNetV2 Bottleneck Block for Coffee Leaf Disease Classification

The deployment of deep learning models for plant disease detection on mobile devices is often limited by the availability of computational resources. Although MobileNetV2 offers an efficient, lightweight architecture, its feature extraction capabilities require enhancement through attention mechanisms without incurring significant computational overhead or relying on underexplored spatial distributions. This study analyzes the effectiveness of strategically distributing attention modules, specifically CBAM, ECA, SimAM, and Coordinate Attention, within MobileNetV2 bottleneck blocks to determine the optimal balance between classification accuracy and computational efficiency. Using a dataset of 1,544 Arabica coffee leaf images classified into Healthy, Phoma , and Cercospora categories, systematic ablation studies were conducted across eight distribution configurations to evaluate performance based on classification and efficiency metrics. This research demonstrated that selective attention distribution significantly outperforms full distribution strategies, where the Center-Focused distribution using Coordinate Attention achieved the highest accuracy of 94.19% with a computational footprint of 2.37M parameters and 302.26 GMACs, while the Edge distribution using ECA proved to be the most balanced configuration, achieving a comparable accuracy of 93.55% with an ultra-fast inference time of 4.67 ms and 299.54 GMACs. These findings confirm that optimizing the spatial distribution of attention modules is more critical than mere architectural depth. Consequently, future research should advance this methodology by integrating Attention Guided Pruning, where attention weights serve as metrics to excise redundant filters for leaner architectures, and exploring Dynamic Neural Architecture Search to determine optimal insertion layers autonomously, rather than relying on static, manual configurations.